In this video presentation from the Linux.conf.au 2015 in Auckland, New Zealand, Douglas Bagnall examines a particularly hot topic in deep learning, namely recurrent neural networks, and all the things you can do with them.
In 2013 Bagnall wrote a Gstreamer plug-in that used a recurrent neural network (RNN) to generate video in imitation of a program it was watching. Pretty soon the same RNN library was being used in another Gstreamer plug-in to classify speech on the radio according to language, and to detect birds by listening for their calls (the language classification is quite accurate and runs at 1500 faster than real time on an old laptop, which is at least a data-point for those wondering about spying capabilities). The RNN has also been used to generate text and code, and to classify text by language and author at a fine-grained level. He shows how the RNN is trained, and how it might be adapted for other forms of time-series data. He demonstrates the various plug-ins and text utilities and, for excitement, execute RNN-generated code on the fly. He also explains what a recurrent neural network is and how it relates to a plain (or “deep”) neural network.
Sign up for the free insideBIGDATA newsletter.
Speak Your Mind